7/31/2019 HotMesh12

1/6

Evaluating the Impact of Frame Aggregation on Video-Streaming overIEEE 802.11n Multihop Networks

Sascha Gbner and Christoph LindemannDepartment of Computer Science

University of LeipzigAugustusplatz 10, 04109 Leipzig, Germany

{guebner, cl}@rvs.informatik.uni-leipzig.de

Abstract In this paper, we study the effects of frameaggregation on video streaming performance in a real-worldIEEE 802.11n testbed. We quantitatively characterize delay,video quality, and mean aggregate size in two multi-hopscenarios. We observe that streaming applications naturallytake advantage of frame aggregation, both in single- and multi-stream environments, and that a Full-HD video can betransmitted with excellent quality and delay, even on a 6-hopchain. Furthermore, we discover that limiting frameaggregation severely impacts both the delay and video quality.We believe these insights on frame aggregation help fordeveloping an effective cross-layer design for video streamingover IEEE 802.11n multi-hop networks.

Keywords: measurements in mesh networks, IEEE 802.11n,video streaming

I. I NTRODUCTION Mobile video is widely considered to become the next

killer application for wireless networks [13]. According to aCisco forecast [6], mobile video traffic will more than doubleevery year till 2015 and has the highest growth rate of anyapplication category measured within the forecast. Thisimposes high demands to the underlying cellular network infrastructure, which possibly cant keep up with this hugetraffic increase. We suppose that multi-hop mesh networksmay be an option to cope the increased traffic demandsarising from mobile video streaming, especially from HighDefinition (HD) video. Other high potential fields of application for video streaming over mesh networks includereal-time video streaming for emergency coordination or wireless video surveillance systems, e.g. for public safety or on building sites.

Although there are many promising scenarios,unfortunately, little is known about the behavior of videostreaming in state-of-the-art mesh networks. To fullyunderstand video streaming in new mesh networks, one mustknow insights on new features of the latest IEEE 802.11standard, namely IEEE 802.11n [12]. With IEEE 802.11n,very high physical data rates up to 600 Mbit/s are achievable.However, these high data rates on the physical layer can only

by harnessed at upper layers, if the medium access isefficient [17]. Therefore, IEEE 802.11n introduces frameaggregation on the MAC layer. With frame aggregation,multiple subframes, each with an own checksum, can be

transmitted in an aggregated frame, with the overhead for medium access arising only once. On reception, the IEEE802.11n MAC can correctly extract individual subframeseven if other parts of the aggregated frame are erroneous.Frame aggregation will also be a key technology in futurestandards, like IEEE 802.11ac [11].

However, the influence of frame aggregation on videostreaming performance in single- and multi-hop wirelessnetworks is not sufficiently investigated. Therefore, we

present in this paper the results of a measurement study onthe impact of frame aggregation on video streaming

performance in a real-world IEEE 802.11n mesh testbed. Weherein focus on two typical scenarios. The first is a singlevideo stream over a chain, e.g. when a mobile user streamsvideo data from the Internet over wireless relay stations. Inthe other scenario, several sources stream and relay to onedestination, as it occurs in wireless video surveillance. Our findings in these scenarios can be summarized as follows:

Streaming applications naturally take advantage of frame aggregation, and the mean aggregate size

increases both with video resolution and path length With IEEE 802.11n, a single Full-HD video can be

transmitted with excellent delay and quality results,even over a 6-hop chain

Limiting the frame aggregation severely impacts both the average delay and the quality of a videostream

In a many-to-one scenario, up to 6 sources canstream an HD video to a common destination withacceptable quality, but favoring the nearest nodes

In the same setup, the mean aggregate size increasesalmost linearly with each new stream

Limiting the aggregation size in this scenario alsoseverely impacts the video quality and average delay

We believe that these insights on frame aggregation arehelpful to develop new enhancements for video streaming inIEEE 802.11n multi-hop networks.

The remainder of this paper is organized as follows.Section II summarizes related work on measurements inIEEE 802.11n networks and video streaming in wirelessmesh networks. In Section III we introduce our experimentation environment; in Section IV themeasurement results are presented. Finally, in Section Vconcluding remarks are given.

7/31/2019 HotMesh12

2/6

II. R ELATED WORK Cheng et al. [5] extensively studied the video streaming

performance in an IEEE 802.11a multi-hop mesh testbed.They conclude that video has its special characteristics andthat higher rates do not always imply higher quality. Dely etal. [8] studied the impact of IP packet aggregation andtransmission opportunities on Voice over IP in an IEEE802.11e testbed. Opposed to [5] and [8], we focus on theimpact of frame aggregation in an IEEE 802.11n testbed.

Chan et al. [4] proposed a video-aware MAC rateadaptation scheme. It adapts frequency and timing of channel

probing to video encoding rate and wireless channelvariations. Jakubczak et al. [13] presented a cross-layer design between application and physical layer. Theymodified the signal transmission on the physical layer to belinearly related to a pixels luminance. Opposed to [4] and[13], we focus on video streaming performance in an IEEE802.11n multi-hop mesh network.

LaCurts et al. [15] analyzed traces gathered fromdifferent real-world wireless mesh networks deployed by

Meraki, using both 802.11b/g and 802.11n devices. Theymainly studied accuracy of SNR-based bit rate adaptation,the impact of opportunistic routing, and the prevalence of hidden terminals. Opposed to [15], we focus on the impact of frame aggregation on video streaming performance in anindoor mesh testbed.

In own previous work [10], we developed an analyticalmodel for characterizing the effective throughput for multi-hop paths in IEEE 802.11n wireless mesh networks as afunction of bit error rate, aggregation level, and path length.In [9], we studied the impact of aggregation and channel

bonding on the achievable throughput on a multi-hop chainin our IEEE 802.11n mesh testbed. Opposed to [9] and [10],we focus in this paper on the impact of frame aggregation on

video streaming performance.In [3], Cai et al. presented an analytical model for studying the impact of frame aggregation and bidirectionaltransmission on voice and video performance for differentaggregation schemes. In [16], Lee et al. detected throughsimulations that multiple-receiver frame aggregation candouble the number of supported video streams. Li et al. [17]

proposed an analytical model to derive the effectivethroughput and optimal frame and fragment sizes for transmissions with frame aggregation. Bononi et al. [2]

proposed a joint channel assignment and multipath routingarchitecture for supporting robust video streaming over multi-radio mesh networks and showed its usability throughsimulations. Opposed to [2], [3], [16], and [17], we

conducted measurements in a real-world IEEE 802.11nmulti-hop testbed.Pefkianakis et al. [18] studied MIMO based rate

adaptation in 802.11n wireless networks and proposed aMIMO aware rate adaptation scheme. Pelechrinis et al. [19],[20] conducted experimental studies on the behavior of MIMO links in different topologies. They mainly focus onthe impact of different 802.11n specific features on the peak

performance [19] and packet delivery ratio under different physical data rates [20]. Opposed to [18], [19], and [20], we

focus on video streaming performance in an IEEE 802.11nmulti-hop mesh network instead of a 1-hop infrastructuremode WLAN.

Deek et al. [7] conducted experimental studies toexamine the influence of channel bonding and interferenceon network performance, in terms of throughput and packetreception probability. Note, that frame aggregation wasdeactivated in most of their experiments, leading to reducedthroughput compared to results in e.g. [9]. Shrivastava et al.[22] studied the impact of channel bonding and interferenceof 802.11g on 802.11n-links in a real testbed deployment.Opposed to [7] and [22], we consider video streaming

performance in multi-hop IEEE 802.11n networks.

III. EXPERIMENTATION E NVIRONMENT

A. Background on IEEE 802.11nThe IEEE 802.11n standard features several

enhancements for higher throughput in IEEE 802.11 wirelesscommunication. Among others, the main features on the

physical layer include channel bonding and the utilization of

MIMO-specific features, like spatial division multiplexing or space-time block coding. Channel bonding allows an IEEE802.11n transmitter to use two adjacent channels in parallel.While data rate increases, the communication becomes moresensible to interference, especially from legacy IEEE802.11a/b/g links, thus collisions and erroneoustransmissions occur more frequently. However, because welimited influence of external interference and to get insightson IEEE 802.11ns full potential, we activated this option.

The other MIMO-specific features are spatial divisionmultiplexing and space-time block coding. Spatial divisionmultiplexing increases the physical data rate by sending twoor more independent data streams in parallel. Space-time

block coding enhances the robustness of a transmission using

a special coding. The choice, whether spatial divisionmultiplexing or space-time block coding is utilized, isencoded in the modulation and coding scheme (MCS) of thetransmitter. For our chipset, which is able to use 2x2 MIMO,MCS classes 0 to 7 use one data stream, while classes 8 to 15use two data streams, encoded with spatial divisionmultiplexing. In [9], we analyzed the distribution of utilizedMCS classes and the limitation of MCS choice onthroughput. Again, to study IEEE 802.11ns full potential,we didnt limit the MCS choice and let the bit rate adaptationalgorithm choose the most appropriate MCS class.

Another important advancement of IEEE 802.11nconcerns the MAC layer. IEEE 802.11n introduces frameaggregation to fully exploit the offered high physical data

rates [17]. Frame aggregation, here MPDU aggregation,allows combining several MAC frames into one larger aggregated frame, with the overhead for medium accessarising only once. Each correctly received frame can beacknowledged by the receiver in a block acknowledgement,reducing the number of frames to be retransmitted. Due tospace limitations, we refer the interested reader to [12]Section 7.4a for details on MPDU aggregation. For our hardware, the maximum amount of frames that can beaggregated is limited to 32 frames.

7/31/2019 HotMesh12

3/6

Figure 1. Indoor MIMO Mesh Testbed

1 2 3 4 5 6 7

Figure 2. Scenario I: Chain topology (one stream with 6 hops)

1 2 3 4 5 6 7

Figure 3. Scenario II: Many-to-one topology (six streams with i hops, i=1, 2, , 6)

B. Indoor MIMO Mesh Testbed Setup

Our Indoor MIMO Mesh Testbed comprises 20 wirelessmesh nodes located in 10 rooms in the department building,covering roughly 250 m. An overview of the testbed withthe node locations is depicted in Figure 1. Each node consistsof a Siemens ESPRIMO P2510 PC with an Intel Celeron 3.2GHz processor, 80 GB HDD, 512 MB RAM, and a D-Link DWA-547 wireless PCI network interface card (NIC). This

NIC is equipped with an AR 9223 Atheros chipset and three5dBi omnidirectional antennas. The AR 9223 is able tosupport MIMO communication in the 2.4 GHz band. Eachnode runs openSUSE 11.2 as operating system with amodified kernel, based on version 2.6.34. Furthermore, weemploy the ATH9K driver for Atheros chipsets.

Each node possesses a Gigabit Ethernet NIC to allow

remote management of the nodes. Hence, wirelessexperiments can be managed from a remote computer, andtraces can be copied and evaluated through the wirednetwork. Table I shows a detailed description of hardwareand software components of the testbed.

C. Experimental setupFor our experimental measurement study, we built up a

6-hop chain topology using the shaded nodes of our indoor MIMO mesh testbed depicted in Figure 1.

TABLE I. T ESTBED OVERVIEW

Component Description

PC Siemens ESPRIMO P2510 Celeron 3.2 GHz,512 MB RAM, 80 GB HDD

Wireless Card D-Link DWA-547 PCI NIC equipped with 3antennas

Chipset Atheros AR 9223, operating at 2.4 GHzOperating System openSUSE 11.2 with kernel version 2.6.34

Note, that we limited the chain length to 6 hops, as we believe its unlikely that longer path lengths will play amajor role in a real-world deployment.

We placed the nodes to let the antennas face into the building to enrich the multi-path scattering, crucial for spatial division multiplexing. All nodes ran in ad-hoc mode,and we assigned static routing between them. To keep thenodes time-synchronized, we setup an NTP daemon on eachnode. Furthermore, we activated channel bonding to achievethe highest possible performance. As bit rate adaptationalgorithm, we used Minstrel HT , the default rate adaptationalgorithm for 802.11n in Linux. Minstrel HT is anadvancement of the widely used SampleRate algorithm [1]

by Bicket.Because our devices use channel bonding and run in the

2.4 GHz band, we conduct the experiments at night or duringthe weekend. In these time slots little interference due to

7/31/2019 HotMesh12

4/6

802.11a/b/g background traffic occurred. We noticed in a previous long-term experiment that during the workinghours, between 8am and 8pm, the measured goodput wasinfluenced by external interference, especially due tostudents who access the Internet wirelessly through their IEEE 802.11 equipped laptops.

For our experiments, we used the video performanceevaluation framework EvalVid [14]. As video test sequences,we extracted the first 60 seconds of the Full-HD videosElephants Dream and Big Buck Bunny from [21]. Note,that although, due to space limitations, only the results of thefirst video are presented, we validated them also with thesecond one. We employed the open source encoder x264 [23] with default settings to create single-layer videosequences, encoded in the widely used video coding standardH.264/MPEG-4 AVC. We encoded each video in differentresolutions according to Table II. We set the Group-of-Picture size to equal the Frames-per-second, which are 24.The H.264 coding profile used was High . With these videosamples, we conducted our streaming experiments startingfrom an initially idle system. We conducted 10 independentreplicates of each experiment and derived the considered

performance measures with 95% confidence level. Thewidths of the confidence intervals are depicted as bars in the

plots.For the actual streaming, we employed the EvalVid

program mp3trace , which transmits a video from the sourceto a given destination, i.e. it uses unicast transmission. Weused UDP as transport protocol to limit the influences of TCPs exponential backoff and retransmit mechanisms. Werun the Linux tool tcpdump on the time-synchronized sourceand destination nodes, to trace delay and loss for each

packet. With these traces, EvalVid is able to calculate delayand loss on a per-video-frame basis and can further reconstruct the video as it would be seen at the receiver. Weuse the sender and receiver side videos to calculate the Peak Signal-to-Noise Ratio (PSNR). PSNR is a standard objectivemetric to measure the quality of an encoded or transmittedvideo compared to the original one. Values around 40dB areconsidered as high quality video without any observabledefect, while videos with a PSNR of 25dB to 30dB couldstill be acceptable, although visible artifacts exists [4].

We chose two scenarios for our experiments as examplesfor the application of video streaming. Figure 2 depictsScenario I that considers a single video stream, transmittedover several hops on a chain topology. This is exemplary for e.g. live television streaming over a wireless mesh with anInternet gateway. However, Scenario II, depicted in Figure 3,considers several video streams in parallel. This is, for example, the case in wireless video surveillance, whenseveral video cameras stream their data in a meshed manner to a central data collection and processing entity.

TABLE II. O VERVIEW OF USED VIDEO SAMPLES

Name Resolution Coding Level Average Rate

[kbit/s] QVGA 320240 1.3 315VGA 640x480 3 960 HD 1280x720 3.1 2376

Full-HD 1920x1080 4 5510

IV. MEASUREMENT RESULTS As an initial experiment, we stream all video samples

over one hop with default configuration. We plot the sendingrate for the 4 videos in Figure 4. We observe that thedifferent videos share several rate peaks, occurring from fastchanging and less compressible scenes, rising up to 13Mbit/s for the Full-HD sample. In the next step, we graduallyincrease the path length up to 6 hops and calculated theaverage delay and PSNR of each video frame. All videos hadexcellent values; therefore, we neglect the correspondingcurves. Instead, we concentrate on the most challengingvideo sample, the Full-HD one. We wanted to study theimpact of frame aggregation on video streaming

performance, in terms of delay and PSNR.Therefore, we vary the maximum allowed number of

MAC frames per aggregate. We denote this as the maximumaggregate size. We plot the results in Figure 5 and Figure 6.We can observe in both figures that decreasing theaggregation severely impacts the streaming quality. In Figure5, we note that for all maximum aggregate sizes the delay

rises almost linearly till a certain hop count, where itsuddenly grows faster. For maximum aggregate size 2, this isalready at 3 hops, while its at 4 hops for maximumaggregate size 4 and 5 hops for larger sizes. We further notethat the increase is slighter for larger maximum aggregatesizes. This effect is due to the increased load and mediumcontention when, on the one hand, path length increases and,on the other hand, maximum aggregate size decreases. Thelatter causes more small aggregates to be transmitted, thusincreasing the overall time when the medium is blocked. InFigure 6, we observe the same effect. With decreasingmaximum aggregate size, the achievable effective capacityof the medium decreases too. This leads to higher loss

probabilities and, therefore, a lower PSNR. We further note

that for a maximum aggregate size of 32 a Full-HD videocan be transmitted without noteworthy distortion, even over a 6 hop chain. Note, that we neglected the confidenceintervals here for clarity, as they overlapped confusingly.

Now we want to have a closer look at the aggregationitself. Figure 7 depicts the mean aggregate size on the lasthop, i.e. the 6th node in Figure 2. We notice that withincreasing resolution also the mean aggregate size rises.

Figure 4. Sending rate vs. time for different resolutions in Scenario I

7/31/2019 HotMesh12

5/6

Figure 5. Average end-to-end delay vs. hops of a Full-HD video for

different maximum allowed aggregate sizes in Scenario I

Figure 6. Average PSNR vs. hops of a Full-HD video for differentmaximum allowed aggregate sizes in Scenario I

Figure 7. Mean aggregate size on last hop vs. hopsfor different resolutions in Scenario I

Figure 8. Fraction of lost video frames vs. path length on a 6-hop chainfor different resolutions in Scenario II

Figure 9. Average PSNR vs. number of parallel streams of an HD videofor different maximum allowed aggregate sizes in Scenario II

Figure 10. Mean aggregate size on last hop vs. number of parallel streamsfor different resolutions in Scenario II

7/31/2019 HotMesh12

6/6

This is evidence that frame aggregation automaticallyadapts to higher traffic load, i.e. with higher mediumcontention, the packet queue increases leading to higher aggregation opportunities. This is also why the meanaggregate size increases slightly with the path length.

The next experiment is conducted for Scenario II.Here, we stream several videos in parallel, i.e. each nodein the chain streams a video to the destination node. Note,that we start the video sequences with some delay to avoidcumulating peaks in the sending rate. This could be thecase, if several surveillance cameras in a rowconsecutively record a moving object. In Figure 8, we firsthave a closer look on the loss each video stream suffers inthe full 6-hop scenario, i.e. 6 nodes stream in parallel tothe destination node. We observe that video streams fromnodes farther to the destination suffer higher losses. For the Full-HD sample, these losses become rapidly risingover 50%. We suppose this is due to queue drops atrelaying nodes that cant access the medium quick enough.

Now we look at the impact of frame aggregation on theaverage Peak Signal-to-Noise Ratio, depicted in Figure 9.We observe that it generally decreases linearly with thenumber of parallel video streams, like in Figure 6. Note,that the delay curve, not shown here, is comparable toFigure 5. We left it due to space limitations. Furthermore,we notice in Figure 9 that, with full frame aggregationenabled, we can stream several HD-videos in parallel withacceptable video quality, even over 6 hops.

In Figure 10, we again vary the chain length but look now closer on the mean aggregate size on the last hop. Weobserve that the mean aggregate size almost proportionallyincreases with the number of video streams that traversethe node. This indicates that frame aggregation helps toalleviate the increasing traffic load, by enlarging thetransmitted physical data units, thus minimizing the time,when the medium is blocked.

V. CONCLUSION We analyzed the impact of IEEE 802.11n frame

aggregation on video streaming performance, in terms of delay and video quality. We further revealed details on thedependency of number of video streams, video resolution,and path length on the aggregation level. We showed that aFull-HD video can be transmitted with excellent delay andquality, even over a 6-hop chain, and that IEEE 802.11n isalso able to support several HD videos in a many-to-onescenario. We further showed that limiting the frameaggregation severely impacts both the average delay andthe quality of a video stream, in both considered scenarios.

For future work, we will use our findings to optimizethe alignment of video transmission and frame aggregationto increase the number of supported videos with 802.11n.

R EFERENCES [1] J. Bicket, Bit-rate Selection in Wireless Networks, Masters Thesis,

MIT, Cambridge, USA, 2005.[2] L. Bononi, M. Di Felice, A. Molinaro, and S. Pizzi, A Cross-Layer

Architecture for Robust Video Streaming over Multi-Radio Multi-

Channel Wireless Mesh Networks, Proc. ACM MobiWac , Tenerife,Canary Islands, 2009.

[3] L. Cai, X. Ling, X. Shen, J. W. Mark, and L. Cai, Supporting voiceand video applications over IEEE 802.11n WLANs, Springer Wireless Networks , 15 , 2009.

[4] A. Chan, H. Lundgren, and T. Salonidis, Video-Aware RateAdaptation for MIMO WLANs, Proc. IEEE ICNP , Vancouver,Canada, 2011.

[5] X. Cheng, P. Mohapatra, S.-J. Lee, and S. Banerjee, Performanceevaluation of video streaming in multihop wireless mesh networks,

Proc. ACM NOSSDAV , Braunschweig, Germany, 2008.[6] Cisco Visual Networking Index: Global Mobile Data Traffic

Forecast Update, 20102015, http://www.cisco.com/en/US/solutions/collateral/ns341/ns525/ns537/ns705/ns827/white_paper_ c11-520862.html, 2011.

[7] L. Deek, E. Garcia-Villegas, E. Belding, S.-J. Lee, and K.Almeroth, The impact of channel bonding on 802.11n network management, Proc. ACM CoNEXT , Tokyo, Japan, 2011.

[8] P. Dely, A. Kassler, N. Bayer, and D. Sivchenko, An ExperimentalComparison of Burst Packet Transmission Schemes in IEEE802.11-based Wireless Mesh Networks, Proc. IEEE GLOBECOM ,Miama, USA, 2010.

[9] J. Friedrich, S. Frohn, S. Gbner, and C. Lindemann,Understanding IEEE 802.11n Multi-hop Communication inWireless Networks, Proc. IEEE WinMee , Princeton, USA, 2011.

[10] S. Frohn, S. Gbner, and C. Lindemann, Analyzing the EffectiveThroughput in Multi-Hop IEEE 802.11n Networks, Elsevier Computer Communications , 34 , 2011.

[11] IEEE P802.11 - TASK GROUP AC, Specification Framework for TGac, IEEE, 2011.

[12] IEEE 802.11n: Standard for Wireless LAN Medium AccessControl (MAC) and Physical Layer (PHY) SpecificationsAmendment 5: Enhancements for Higher Throughput, IEEE, 2009.

[13] S. Jakubczak and D. Katabi, A cross-layer design for scalablemobile video, Proc. ACM MOBICOM , Las Vegas, USA, 2011.

[14] J. Klaue, B. Rathke, and A. Wolisz, EvalVid - A Framework for Video Transmission and Quality Evaluation, Proc. TOOLS ,Urbana, USA, 2003.

[15] K. LaCurts and H. Balakrishnan, Measurement and Analysis of Real-World 802.11 Mesh Networks, Proc. ACM IMC , Melbourne,Australia, 2010.

[16] K. Lee, S. Yun, and H. Kim, Boosting Video Capacity of IEEE802.11n through Multiple Receiver Frame Aggregation, Proc.

IEEE VTC , Marina Bay, Singapore, 2008[17] T. Li, Q. Ni, D. Malone, D. Leith, Y. Xiao, and R. Turletti,

Aggregation with Fragment Retransmission for Very High-SpeedWLANs, IEEE/ACM Transactions on Networking , 17 , 2009.

[18] I. Pefkianakis, Y. Hu, S. H. Wong, H. Yang, and S. Lu, MIMORate Adaptation in 802.11n Wireless Networks, Proc. ACM

MOBICOM , Chicago, USA, 2010.[19] K. Pelechrinis, T. Salonidis, H. Lundgren, and N. Vaidya,

Analyzing 802.11n Performance Gains, Proc. ACM MOBICOM (poster session), Beijing, China, 2009.

[20] K. Pelechrinis, T. Salonidis, H. Lundgren, and N. Vaidya,

Experimental characterization of 802.11n link quality at high rates, Proc. ACM WINTECH , Chicago, USA, 2010.[21] P. Seeling and M. Reisslein, Video Transport Evaluation With

H.264 Video Traces, IEEE Communications Surveys & Tutorials ,to appear

[22] V. Shrivastava, S. Rayanchu, J. Yoon, and S. Banerjee, 802.11nUnder the Microscope, Proc. ACM IMC , Vouliagmeni, Greece,2008.

[23] X264, H.264/AVC Encoder, http://www.videolan.org/developers/x264.html